Rather than doing this lowering potentially multiple times when a
shader is relinked we can instead do it once in the compiler.
This change also gets us closer to converting to NIR at compile
time.
Reviewed-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27690>
We have competent lowering in NIR already available.
Drivers exposing CAP_DOUBLES but not SHADER_CAP_DROUND:
- d3d12 (NIR lowers ~0 if the underlying impl doesn't do floats)
- svga (Now sets the NIR lowering options)
- softpipe (Doesn't do GL4 so you can't use doubles anyway)
- llvmpipe (Lowers dround_even in NIR and passees the rest through
successfully)
- zink (NIR lowers ~0 if the underlying impl doesn't do floats,
otherwise passes things through successfully, except needed
dround_even lowering to avoid lavapipe regression with
native doubles)
- r600 (sets NIR rounding lowering flags, and lowers all fsign)
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Acked-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25777>
Most drivers that can expose GL4 were claiming the cap anyway (llvmpipe,
softpipe, zink, iris, nvc0, radeonsi, r600, freedreno, d3d12), and just
doing lowering in NIR if nessary.
crocus was only claiming the cap for gen8, but the backend compiler
enables NIR lowering regardless.
svga is the only other GL4 driver that didn't set it, and we can just set
the NIR lowering flag.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Acked-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25777>
Replaced by the new NIR pass.
i915g results:
total instructions in shared programs: 510678 -> 510714 (<.01%)
total temps in shared programs: 30429 -> 30426 (<.01%)
rv370 results:
total instructions in shared programs: 737649 -> 737656 (<.01%)
instructions in affected programs: 82 -> 89 (8.54%)
total temps in shared programs: 112093 -> 112094 (<.01%)
temps in affected programs: 6 -> 7 (16.67%)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24763>
All drivers should now be using the appropriate NIR lowering, so we can
drop this pile of code.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22083>
NIR i/o lowering and sysval lowering can handle the compact var fine at
this point.
Affects: nouveau, virgl, svga, radeonsi, r600, llvmpipe. Does not affect
PIPE_CAP_NIR_COMPACT_ARRAYS drivers like crocus, iris, d3d12, freedreno,
zink.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21940>
Now that we don't do GLSL IR constant propagation or constant folding, we
can leave constant variable detection and handling to NIR. This also
avoids some OOB array access in GLSL IR in a piglit test!
Freedreno stats again look like noise:
total instructions in shared programs: 2718412 -> 2718746 (0.01%)
instructions in affected programs: 80497 -> 80831 (0.41%)
total last-baryf in shared programs: 110015 -> 110510 (0.45%)
last-baryf in affected programs: 35263 -> 35758 (1.40%)
total full in shared programs: 189486 -> 189480 (<.01%)
full in affected programs: 52 -> 46 (-11.54%)
total constlen in shared programs: 494540 -> 494496 (<.01%)
constlen in affected programs: 452 -> 408 (-9.73%)
total sstall in shared programs: 198297 -> 197928 (-0.19%)
sstall in affected programs: 3691 -> 3322 (-10.00%)
total systall in shared programs: 432150 -> 431799 (-0.08%)
systall in affected programs: 6070 -> 5719 (-5.78%)
total waves in shared programs: 435098 -> 435110 (<.01%)
waves in affected programs: 92 -> 104 (13.04%)
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21751>
NIR is happy to take care of constant folding for us, and it's easier to
do in SSA.
This required adjusting of lower_precision unit tests to have un-folded
constants.
freedreno results look like noise. Some excerpts:
total instructions in shared programs: 2718343 -> 2718412 (<.01%)
instructions in affected programs: 6847 -> 6916 (1.01%)
total last-baryf in shared programs: 109992 -> 110015 (0.02%)
last-baryf in affected programs: 117 -> 140 (19.66%)
total sstall in shared programs: 198312 -> 198297 (<.01%)
sstall in affected programs: 148 -> 133 (-10.14%)
total systall in shared programs: 432163 -> 432150 (<.01%)
systall in affected programs: 1016 -> 1003 (-1.28%)
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21751>
We already have a nir_builder equivalent for generating this code, just
use that instead of doing it in GLSL.
No change on r300 shader-db.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21476>
GLSL's lower_vector_derefs already does this, and even if it didn't
nir_vector_extract() would when glsl-to-nir happens.
No effect on freedreno shader-db.
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18466>
nir_lower_vars_to_ssa will split temp structs up anyway. This fixes a bug
where mediump wouldn't be propagated to the split vars.
The effect is tiny, I think just shuffling some code scheduling from
optimizing at different places. Affects Natural Selection 2, Serious Sam
3, 3dmark slingshot, and Lego Legacy.
freedreno shader-db:
total instructions in shared programs: 11315637 -> 11315993 (<.01%)
instructions in affected programs: 24861 -> 25217 (1.43%)
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18466>
nir_lower_vars_to_ssa will split temp arrays up anyway. Fixes a bug where
split arrays wouldn't get their precision qualifier.
Helps mostly Android and skia shaders. Also affects Civ5, Witcher 2, and
Borderlands 2.
freedreno shader-db:
total instructions in shared programs: 11319395 -> 11319355 (<.01%)
instructions in affected programs: 65744 -> 65704 (-0.06%)
Reviewed-by: Adam Jackson <ajax@redhat.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18466>
We have fine NIR lowering for this (already called from mesa/st), no need
for a separate GLSL pass.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18361>
Every driver uses the nir_lower_system_values path now.
Reviewed-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18327>
The nir_opt_conditional_discard pass is called anyway and covers
discard/demote/terminate.
iris shader-db:
total instructions in shared programs: 8933422 -> 8933426 (<.01%)
instructions in affected programs: 48 -> 52 (8.33%)
helped: 0
HURT: 4
which is a synmark shader going from 12 to 13 instrs.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17664>
It will get turned into SSA and copy-propagated in NIR, no need to walk
the IR collapsing it here.
iris shader-db results appear to be noise:
total instructions in shared programs: 8932195 -> 8932147 (<.01%)
instructions in affected programs: 537 -> 489 (-8.94%)
LOST: 12
GAINED: 11
lost/gained are simd32 switches in unigine, l4d2, portal2, asphalt9.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17613>
Now that we have no non-NIR drivers, we can retire the old code. We just
need to pass the variable accesses through to it.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17610>
Since we have now switched all drivers to using NIR and therefore
the NIR based uniform linker this param never needs to be set to
true so remove it.
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16924>
nir_lower_int_to_float() does this at the end of compilation, no need to
do it up front.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16823>
NIR has fdiv, and all the NIR backends have to have lower_fdiv set
appropriately already since various passes (format conversions,
tgsi_to_nir, nir_fast_normalize(), etc.) might generate one.
This causes softpipe and llvmpipe to now do actual divides, since
lower_fdiv is not set there. Note that llvmpipe's rcp implementation is a
divide of 1.0 by x, so now we're going to be just doing div(x, y) instead
of mul(x, div(1.0, y)).
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16823>
It's way more concise to write as nir_builder calls.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16823>
The drivers not setting it were:
- nv30, which gets lowering using NIR's lower_fsat flag.
- r300, which gets lowering using NIR's lower_fsat flag.
- a2xx, which has was getting it optimized back to fsat anyway.
This drops the check for the cap from gallium nine. While nine does have
a non-nir path, I think it's safe to assume that if you have SM3
texturing, you can do fsat.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16823>
NIR already has the necessary lowering, and the GLSL lowering violates
GLSL IR validation rules. Once quadop lowering was turned off, the IR
validation at the end of the compile path on DEBUG builds caught the
problem.
In order to move the lowering to NIR, though, we need to make sure that
drivers supporting these functions actually have the lowering flag set.
xfails added for t860, where apparently this tickles a variety of existing
64-bit bugs in the backend.
Fixes: #6461
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Mykhailo Skorokhodov <mykhailo.skorokhodov@globallogic.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16437>
The single caller of this function (in st_glsl_to_ir.cpp) always passes
false, so this is dead code.
v2: Delete convert_vec_index_to_cond_assign method because all the
callers are deleted too.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16440>
As of ca63a5ed3e ("glsl: fix interpolateAtXxx(some_vec[idx], ...) with
dynamic idx"), this lowering pass does two things. It converts
ir_binop_vector_extract to an if-ladder to select the dynamically
indexed component, and it extracts a ir_binop_vector_extract from the
source of an interpolateAt function and applies to the result instead.
This change adds a flag to disable the former behavior. The latter is
still useful, but NIR has better (and soon even better) ways of doing
the former.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16440>
Now that everybody goes through NIR, glsl_to_nir is happy to handle the
instruction and turn it into nir_op_vec4 instead of going to a temp
variable and back.
No changes on freedreno shader-db.
Reviewed-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16363>
Now that everything goes through NIR, nir_opt_peephole_select has replaced
it.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8044>
All UBO-supporting drivers now go through the NIR path, which does a
better job of it.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8044>
